Personal Agentic AI Operating System

by Artem Kroupenev · Substack

A battle-tested foundation for building a personal agentic AI OS on OpenClaw.

Built from a real production setup running daily — 14 agents, 15+ crons, hundreds of sessions worth of iteration. Not a toy. This is the architecture that actually worked.

---

What This Is

A complete workspace template that gives you:

• Multi-agent pipeline — creator agents produce work, critic agents stress-test it before it reaches you
• Persistent memory — 4-layer memory stack that compounds over time (semantic search, daily notes, structured learnings, live state)
• Research methodology — Deep Research Protocol (3-layer insight funnel) + Atom of Thought (structured reasoning with uncertainty disclosure)
• Security built in — hard rules for email, money, git, and irreversible actions. PIN guards. Prompt injection defense.
• Brain vault — Obsidian-compatible knowledge base for people, projects, topics, and your writing voice
• Task architecture — single task queue with goal ancestry, atomic checkout, and engagement tagging

---

Quick Start

1. Install OpenClaw

npm install -g openclaw
openclaw init

2. Clone this kit into your workspace

git clone https://github.com/YOUR_USERNAME/grace-openclaw-kit.git
# Point your OpenClaw workspace at the cloned folder

3. Fill in USER.md (5 minutes)

4. Tell your agent:

"Read SETUP.md and help me set up my OpenClaw instance"

The agent walks through setup interactively — personalizes your SOUL.md, installs recommended skills, sets up crons.

---

What's In Here

grace-openclaw-kit/
├── SETUP.md              ← Agent reads this to run setup
├── USER.md               ← Fill in: who you are
├── SOUL.md               ← Your agent's identity (template)
├── AGENTS.md             ← Session protocols + context-first rules
├── SECURITY.md           ← Hard rules — email, money, git, privacy
├── MEMORY.md             ← Long-term semantic memory (grows over time)
├── TOOLS-AND-APPS.md     ← What to install and connect
│
├── agents/
│   ├── execution-protocol.md         ← 🟢🟡🔴 complexity gate
│   ├── creator-critic-pattern.md     ← Creator → Critic pipeline + grade system
│   ├── process-validator-template.md ← HOW auditor (Validator tier)
│   ├── agent-registry-template.yaml  ← Pipeline map with routing rules
│   ├── deep-research-protocol.md     ← 3-layer insight funnel methodology
│   ├── atom-of-thought.md            ← Structured reasoning with uncertainty
│   ├── edge-strategies.md            ← Context flow between agents
│   ├── network-creator-template.md   ← Relationship management agent
│   ├── reflection-protocol.md        ← Post-session learning extraction
│   ├── consolidation-protocol.md     ← Monthly memory maintenance
│   ├── session-end-protocol.md       ← Multi-agent context handoff
│   ├── compaction-protection.md      ← Context window survival
│   └── grading-protocol.md          ← A-/B+ quality gates
│
├── state/
│   ├── hot-context.md         ← Live grounding + compaction recovery
│   ├── checkpoint.md          ← Current status + open items
│   ├── tasks.yaml             ← Task queue (single source of truth)
│   ├── goals.yaml             ← Mission → goal → task hierarchy
│   ├── project-memory.yaml    ← Filtered persistent learnings
│   ├── system-overview.yaml   ← Current system state
│   ├── evals.yaml             ← Validator grade log
│   ├── activity-feed.yaml     ← Agent action audit trail
│   ├── contacts.yaml          ← Network relationship database
│   └── top-patterns.md        ← Cross-agent pattern aggregation
│
├── memory/
│   ├── notes.md               ← Append-only observations
│   ├── decisions.md           ← Key decisions + rationale
│   └── YYYY-MM-DD-template.md ← Daily notes template
│
└── data/brain/
    ├── README.md              ← How to use the brain vault
    ├── content/               ← Voice DNA + writing style
    ├── people/                ← Contact knowledge files
    ├── projects/              ← Active work files
    └── topics/                ← Concepts and frameworks

---

Key Principles

Memory that compounds. The brain vault + 4-layer memory stack means your agent gets smarter over time. Every important fact has a source. Every lesson passes a 4-test quality filter before it's stored.

Security is explicit. Hard rules prevent the most common agent mistakes — sending emails without review, auto-accepting financial actions, force-pushing git. These aren't suggestions; they're enforced at the identity layer.

Creator → Critic → Validator pipeline. Every creator has an adversarial critic. Hard A- ship threshold. 3-round iteration cap, then escalates to you. Process validator checks HOW agents work; domain critics check WHAT they produced.

Research methodology built in. Deep Research Protocol prevents lazy AI research (load real artifacts, 3-layer insight funnel, stress test). Atom of Thought prevents confident hallucination (decompose → validate → label → disclose uncertainty).

Tasks as source of truth. Everything open lives in tasks.yaml. Agents don't hold context between sessions — they read the file. Nothing gets lost.

---

Recommended Stack

Tool	Purpose
OpenClaw	Runtime
Claude Code	Coding agent
Obsidian	Brain vault UI (open data/brain/ as vault)
gog (clawhub)	Google Workspace — Gmail, Calendar, Drive
gh (brew)	GitHub CLI
1Password CLI	Secret management
Discord or Telegram	Agent communication channel

Full install guide: TOOLS-AND-APPS.md

---

Lessons Baked In

This kit encodes real lessons from production use:

• AGENTS.md kept under 15K chars — OpenClaw silently truncates at 20K. Extract large sections to separate files that agents Read on demand.
• Memory search tuned — MMR (lambda 0.7) + temporal decay (30-day half-life) reduce duplicate snippets and favor recent context.
• Cron timeouts sized for thinking — If using thinking: high, budget 300-600s timeouts. Sonnet handles most cron workloads fine without extended thinking.
• PIN at the identity layer — Security rules in shared files don't propagate to creator agents. Put PIN rules in each agent's SOUL.md, not a shared SECURITY.md.
• Daily notes during sessions, not after — Don't wait for a nightly reflection cron. Write raw material during the session so the cron has something to distill.
• Verify before claiming — Agents hallucinate infrastructure facts confidently. Always check the actual config/file before stating "you already have X configured."

---

Post-Setup Recommendations

After your first week:

1. Run openclaw doctor — catches config drift and warns about bootstrap file size limits
2. Set up nightly git backup — cron that commits and pushes workspace changes
3. Enable memory search — configure embedding provider in openclaw.json for semantic recall
4. Create a reflection cron — nightly learning extraction compounds fast

---

⚠️ Privacy Notice

Once you fill in USER.md, SOUL.md, MEMORY.md, and daily notes with personal data, keep this repo private. The .gitignore excludes the most sensitive files, but always review before any push.

---

License

MIT — see LICENSE